Transfer learning for EEG-based BCIs: a comparative evaluation and optimization of data alignment methods

Area of Science:

• Neuroscience
• Computer Science
• Biomedical Engineering

Background:

• Practical Electroencephalography (EEG)-based brain-computer interfaces (BCIs) face challenges in cross-subject generalization due to individual brain signal differences.
• Leveraging existing subject data to improve performance for new users with minimal calibration is a critical need.

Purpose of the Study:

• To systematically compare and optimize data alignment techniques for EEG-based BCIs.
• To enhance cross-subject generalization by mitigating individual differences in brain signals.

Main Methods:

• Comparison of Riemannian Procrustes Analysis (RPA), Euclidean Alignment (EA), and Correlation Alignment (CORAL) for transforming EEG data into a common space.
• Leave-one-subject-out cross-validation (LOSO-CV) on EEG attention decoding data.
• Optimization of key parameters, specifically the regularization parameter α for EA.

Main Results:

• Alignment methods improved classification accuracy compared to a no-alignment baseline.
• Optimized EA (at α = 100) yielded the largest mean improvement, increasing accuracy by 3.44%.
• Subject-specific differences in optimal alignment strategies were observed.

Conclusions:

• Optimized alignment techniques, particularly EA, can significantly enhance cross-subject transfer learning in EEG-BCIs.
• This research provides a framework for quantifying alignment benefits and highlights the value of parameter optimization.
• Results pave the way for more robust and generalizable BCI systems with reduced calibration needs.